MCPcopy Create free account
hub / github.com/MeiGen-AI/MultiTalk / unpatchify

Method unpatchify

wan/modules/multitalk_model.py:776–799  ·  view source on GitHub ↗

r""" Reconstruct video tensors from patch embeddings. Args: x (List[Tensor]): List of patchified features, each with shape [L, C_out * prod(patch_size)] grid_sizes (Tensor): Original spatial-temporal grid dimensions before patc

(self, x, grid_sizes)

Source from the content-addressed store, hash-verified

774
775
776 def unpatchify(self, x, grid_sizes):
777 r"""
778 Reconstruct video tensors from patch embeddings.
779
780 Args:
781 x (List[Tensor]):
782 List of patchified features, each with shape [L, C_out * prod(patch_size)]
783 grid_sizes (Tensor):
784 Original spatial-temporal grid dimensions before patching,
785 shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches)
786
787 Returns:
788 List[Tensor]:
789 Reconstructed video tensors with shape [C_out, F, H / 8, W / 8]
790 """
791
792 c = self.out_dim
793 out = []
794 for u, v in zip(x, grid_sizes.tolist()):
795 u = u[:math.prod(v)].view(*v, *self.patch_size, c)
796 u = torch.einsum('fhwpqrc->cfphqwr', u)
797 u = u.reshape(c, *[i * j for i, j in zip(v, self.patch_size)])
798 out.append(u)
799 return out
800
801 def init_weights(self):
802 r"""

Callers 1

forwardMethod · 0.95

Calls

no outgoing calls

Tested by

no test coverage detected